feat(gpu-fleet): roster-driven GPU config + Ornith-1.0-35B

2026-06-28 16:24:54 +00:00
parent 8e0f6e407b
commit e9aa220ee8
3 changed files with 140 additions and 15 deletions
@@ -0,0 +1,75 @@
+# GPU Roster Loader - reads gpu_roster.yaml via PyYAML
+# Hot-reloadable via /admin/roster/reload endpoint
+
+import os, json, threading, time
+
+try:
+    import yaml
+except ImportError:
+    yaml = None
+
+ROSTER_PATH = os.environ.get('ROSTER_PATH', '/app/gpu_roster.yaml')
+_last_mtime = 0
+_lock = threading.Lock()
+
+GPU_URLS = {}
+GPU_SIDECARS = {}
+GPU_LABELS = {}
+GPU_MAX_CONCURRENT = {}
+GPU_CONTEXT = {}
+TIER_MODELS = {}
+HOSTS = {}
+
+def load_roster(path=None):
+    global GPU_URLS, GPU_SIDECARS, GPU_LABELS, GPU_MAX_CONCURRENT
+    global GPU_CONTEXT, TIER_MODELS, HOSTS, _last_mtime
+    if yaml is None:
+        return False, 'PyYAML not installed - run: pip3 install pyyaml'
+    path = path or ROSTER_PATH
+    try:
+        with open(path) as f:
+            data = yaml.safe_load(f)
+        with _lock:
+            GPU_URLS.clear()
+            GPU_SIDECARS.clear()
+            GPU_LABELS.clear()
+            GPU_MAX_CONCURRENT.clear()
+            GPU_CONTEXT.clear()
+            TIER_MODELS.clear()
+            models = data.get('models', {})
+            for name, cfg in models.items():
+                GPU_URLS[name] = cfg.get('gpu_url', '')
+                GPU_SIDECARS[name] = cfg.get('sidecar_url', '')
+                GPU_LABELS[name] = cfg.get('label', name)
+                GPU_MAX_CONCURRENT[name] = cfg.get('max_concurrent', 1)
+                GPU_CONTEXT[name] = cfg.get('context', 65536)
+                tiers = cfg.get('tiers', ['enterprise'])
+                for t in tiers:
+                    if t not in TIER_MODELS:
+                        TIER_MODELS[t] = []
+                    if name not in TIER_MODELS[t]:
+                        TIER_MODELS[t].append(name)
+            HOSTS.clear()
+            HOSTS.update(data.get('hosts', {}))
+            _last_mtime = os.path.getmtime(path)
+        return True, 'Loaded {} models, {} hosts'.format(len(GPU_URLS), len(HOSTS))
+    except Exception as e:
+        return False, str(e)
+
+def check_reload():
+    global _last_mtime
+    try:
+        mtime = os.path.getmtime(ROSTER_PATH)
+        if mtime > _last_mtime:
+            success, msg = load_roster()
+            if success:
+                print('[ROSTER] Auto-reloaded:', msg)
+    except:
+        pass
+
+def reload_thread(interval=30):
+    while True:
+        time.sleep(interval)
+        check_reload()
+
+threading.Thread(target=reload_thread, daemon=True).start()